Astraeus: An Enterprise Financial Forensic Audit Automation Platform Based on Multi-Agent Architecture

Astraeus is a production-grade multi-agent orchestration platform designed to address challenges in traditional manual financial audits (such as the difficulty of identifying factual inconsistencies between SEC 10-K annual reports and earnings call transcripts). Its core innovation is the Lead Auditor-Critic architecture, which enables automated enterprise financial forensic audits through collaboration among specialized AI agents. It provides a complete practical reference from architectural design to production deployment, demonstrating the application potential of multi-agent systems in complex business scenarios.

Traditional manual financial audits face the challenge of identifying factual inconsistencies between SEC 10-K annual reports and earnings call transcripts, requiring significant professional knowledge and time investment. To address this issue, the Astraeus project was born—it is a production-grade multi-agent orchestration platform specifically for automating enterprise financial forensic audits. Its core innovation lies in the Lead Auditor-Critic architecture, which automatically detects discrepancies between official filing documents and management's oral statements through multi-agent collaboration.

Astraeus uses LangGraph to build a state-aware directed graph execution engine, modeling the audit process as state transitions between nodes. The system includes multiple specialized agents:

• Request Gatekeeper: Verifies query security and scope, and performs system health checks;
• The Planner: Breaks down user requests into subtasks, classifying them into quantitative analysis (Type A), qualitative theme analysis (Type B), and discrepancy audit (Type C);
• The Retriever: Performs similarity searches based on the Qdrant vector database and dynamically pulls relevant document fragments;
• The Critic: Verifies the accuracy of retrieved documents, triggers feedback loops, or saves evidence to the audit wiki;
• Unified Generator: Integrates evidence to generate professional audit reports;
• Audit Engine: Performs in-depth verification, calculating metrics such as hallucination scores and mathematical accuracy.

Data Pipeline: Uses DVC for data version management. The process includes multi-source data ingestion (S3/local PDFs), structured extraction (text/tables), PII desensitization (Microsoft Presidio), semantic chunking, and metadata tagging (to ensure data accuracy). Observability:

• LangSmith full-link tracing to visualize agent workflows;
• Prometheus monitoring for end-to-end latency (baseline 53.11 seconds) and node performance;
• Memory guard mechanism to prevent overflow;
• MLflow records token consumption, costs, and traces to support traceability.

Astraeus reduced the total audit time from 5-6 minutes to 53 seconds through the following optimizations:

1. Breakthrough of Retriever-Critic Bottleneck: Pre-filtering layer prunes unnecessary data, reducing latency from 240 seconds to 19.45 seconds;
2. Audit Wiki: Persists short-term memory, skips redundant retrieval tasks, and achieves instant responses;
3. Evidence Summary Delivery: Only passes verified evidence summaries to the generator, controlling the context window (average 3596 tokens) to reduce costs and pressure.

Audit Types:

• Type A (Quantitative Analysis): Calculates financial metrics (e.g., gross profit margin, changes in cash and cash equivalents);
• Type B (Qualitative Theme Analysis): Analyzes management discussion content (e.g., digital sales growth);
• Type C (Discrepancy Audit): Identifies inconsistencies between 10-K reports and earnings call records (e.g., discrepancies between digital acceleration discussions and revenue lines). Quality Assessment: Uses the RAGAS framework, with a faithfulness score of approximately 88% (ensuring zero data fabrication) and an answer relevance score of approximately 75% (Type C still needs optimization).

Astraeus represents cutting-edge practice of multi-agent systems in enterprise-level applications, with core value in transforming AI capabilities into deployable, monitorable, and trustworthy production systems. Insights for developers:

1. State-aware multi-agent architecture can handle complex business processes;
2. Technologies like pre-filtering and intelligent caching improve performance;
3. Production-grade AI requires full-link monitoring and evaluation;
4. Introduce human review at key decision points to balance automation and reliability. This open-source project provides a reference implementation for fields such as financial auditing.